MCP interface: Connecting AI assistants
Updated Reading time: 10 min
On this page
Employees connect external AI assistants such as Claude, ChatGPT, or Copilot to your platform via the MCP interface. The assistant always acts on behalf of the person and with exactly their permissions. Conversely, Kai, the platform's AI assistant, can use tools from external services.
Info
New since version 6.94 (10/22/2026): You now control who can connect AI assistants via the "Connect AI assistants via MCP" permission in roles. By default, only system admins have it. There is no longer a separate toggle for the interface.
Since version 6.93 (10/21/2026): Under Security, you can specify that new AI assistants only get access after admin approval.
Since version 6.92: Each person sees and manages their own connected AI assistants in their profile under Integrations → AI Assistants (MCP). Microsoft Copilot Studio is now recognized as a known client.
Since version 6.91: The MCP Server page is divided into four tabs. Any AI assistant with an HTTPS address can self-register, and the consent page shows where the access is going. Under Security, you can restrict registration to known clients and get notified about new clients.
What is MCP?#
The Model Context Protocol (MCP) is an open standard that allows AI assistants to access other systems. Kibi Connect uses it in two ways:
-
External assistants on your platform
Claude, ChatGPT, Copilot, and other programs read content and act on behalf of a logged-in user—for example, searching wiki pages, creating tasks, or entering appointments.
-
Kai with external tools
Kai uses tools from external MCP servers, such as your ERP or CRM system, n8n, or Notion, without employees having to leave the platform.
Allowing External AI Assistants Access#
The MCP Server in Administration#
Open Administration → Integrations → MCP Servers. Only system admins can view and edit this page. It has four tabs:
-
Overview
Under Who can connect AI assistants?, you see which roles have the "Connect AI assistants via MCP" permission. You control this via roles; by default, only admins have it. Open Roles takes you directly to this permission in the role editor. Below that are the connection URL, which you share with your employees, and a brief explanation of how registration works.
-
Connected Clients
A table of all AI assistants accessing your platform: client name and address, usage (active accesses), and last access. Labels indicate whether a client is Verified: <Address> or uses an Unknown Domain. Clicking a row opens the details with all callback addresses. Use Revoke to end access.
-
Security
Here you specify which clients can newly connect (Allow only known clients), what happens with new AI assistants (Off, Notify only, or Admin approval required), and who gets notified. Under Requests, you see pending and rejected approval requests.
-
Instructions & Technical Details
Pre-written text for employees to copy into their AI chat—the assistant then guides them step-by-step through setup, including for Claude, Claude Code, Langdock, Microsoft Copilot Studio, and ChatGPT. Below that are the technical addresses (endpoints) in case a program asks for them.
On the "MCP Servers" card under Integrations → Overview, you also see for whom MCP applies: Active for: <Roles> or Not enabled for any role.
Warning
Which tools an assistant can use is not determined by MCP server settings, but by three things: whether a role of the logged-in person has the "Connect AI assistants via MCP" permission, which modules are active on your platform, and what additional permissions the person has. An external assistant never sees or does more than the person themselves is allowed to do on the platform.
Who Can Connect AI Assistants#
Whether a person can connect AI assistants is determined solely by their role. The permission is called "Connect AI assistants via MCP" and is found in the role editor under Administration → Roles & Permissions in the System tab, under the AI Assistants (MCP) section.
-
Default
Only the System Admin role has this permission. To enable it for other roles, activate it in the role editor, for example via Open Roles in the Overview tab.
-
Guests
Guests can never connect AI assistants, even with a role.
-
Without the permission
The person does not see an AI Assistants (MCP) section in their profile. If they still try to connect, the consent page shows: "Connecting AI assistants is not enabled for your account. Contact your administration." No approval request is created for the administration.
-
Permission revoked
If a person loses the permission, the platform immediately blocks their existing connections. If they regain it, the connections work again.
-
Existing platforms
The previous setting has been adopted: Wherever MCP was enabled, System Admins have the permission. You can enable it for other roles in the role editor.
Completely Disable MCP#
To completely disable the interface, remove the "Connect AI assistants via MCP" permission from all roles—exceptionally, including the System Admin role. If no role has this permission, no AI assistant can find the interface: registration and login respond with "not found" (404), and connected assistants lose access.
Warning
Warning when resetting: "Reset to default" restores the permission to the System Admin role. The dialog points this out: If MCP is currently not enabled for any role, it will be enabled again afterward.
Security: Which Assistants Can Connect#
Who can connect is controlled by roles. Under Security, you specify what can connect: Any AI assistant with an HTTPS address can self-register with your platform (Dynamic Client Registration). However, it only gets access when a person with the "Connect AI assistants via MCP" permission logs in and explicitly consents on the consent page. There, they see what the assistant can read and do on their behalf and where the access is going.
Warning
Yellow warning on the consent page: If the application is not a known client, a yellow notice appears with the address where the approval code is sent. Only allow access if you set up the connection yourself.
Assistants that identify themselves via a client metadata document—such as Claude, ChatGPT, and VS Code—appear as Verified, both on the consent page and under Connected Clients.
Allow Only Known Clients#
Under Security, you can restrict new connections to a list of domains. The list is pre-populated with known clients—Claude, ChatGPT, Cursor, Langdock, and Microsoft Copilot Studio—and can be edited: remove entries or add your own domains (only the address with https://, no path, e.g., https://app.example.com). Use Restore default list to revert to the pre-populated list.
-
Assistants on your own computer
Programs like Claude Code that run on the user's computer remain allowed even with the restriction enabled.
-
Existing connections
Enabling the restriction does not revoke anything. Therefore, check under Connected Clients if anything is listed that is not on the list and revoke it if necessary.
New AI Assistants: Notify or Approve#
Under Security, in the New AI Assistants card, you specify what happens when someone connects an AI assistant for the first time:
-
Off
No notification. Each person confirms the connection themselves.
-
Notify only
The administration receives a notification as soon as someone connects an AI assistant for the first time. This way, unwanted access is quickly noticed and can be revoked.
-
Admin approval required
A new AI assistant only gets access after the administration approves it. The person waits until then. Assistants on the list of allowed clients are already approved. As long as "Allow only known clients" is enabled, this option is grayed out—only what is on the list can connect anyway.
Below that, under Who gets notified (and approves)?, you choose either All Admins or Specific Admins—one or more people with the System Administrator role. If none of the selected people are admins anymore, notifications and requests go back to all admins.
How Approval Works#
A person connects an AI assistant whose address is not yet approved. After logging in, they see a waiting page: "Keep this window open"—as soon as approval is granted, it continues automatically.
The selected admins receive a push notification and a card in the chat. Clicking it opens the Check AI Assistant dialog: name, address, whether the assistant is verified or unknown, how many people are already connected, and who requested it. Clicking the name opens a chat with that person.
With Allow, the address is approved: all future connections from the same assistant go through without a new request. With Deny, you can optionally provide a reason that the person sees in their notification.
If the person closed the waiting page, they receive a message: "
-
Approve previously denied requests
Under Security → Requests, you see pending and denied requests. A denied request can be approved retroactively there with Approve Anyway.
-
Assistants on your own computer
Local assistants like Claude Code or programs in the development environment (IDE) do not require approval.
-
Existing connections
If you enable approval, existing connections remain valid.
Connecting Assistants (Guide for Employees)#
There is no API key or password to share. Each person connects their assistant themselves:
Info
Prerequisite: One of your roles has the "Connect AI assistants via MCP" permission. By default, only System Admins have it. If you lack this permission, you won’t see an AI Assistants (MCP) section in your profile—contact your administration in that case.
Enter the connection URL in the AI assistant (see setup per program). Each person finds it in their profile under Integrations → AI Assistants (MCP).
The assistant automatically registers with your platform and opens a browser window with the login page.
The person logs in and confirms on the consent page with "Allow access." The page states what the assistant can read and do on their behalf and where the access is going.
Done: The assistant now operates on behalf of this person. Access is valid for 365 days. The person can disconnect it themselves in their profile, and System Admins can revoke it under Connected Clients.
If the administration has enabled approval for new AI assistants and the assistant is not yet approved, a waiting page appears after login. Keep the window open—once approved, it continues automatically. If you close it, you’ll receive a notification and can simply reconnect the assistant after approval.
Your Assistants in Your Profile#
In your profile, under the Integrations tab, there is a section called AI Assistants (MCP), between "Devices and Apps" and "API." It is only visible if one of your roles has the "Connect AI assistants via MCP" permission. If nothing else appears in the Integrations tab, the tab itself is also missing without the permission. The section contains:
-
Connection URL
The address to enter in the AI assistant, with a copy button, and a brief explanation of how the connection works.
-
Instructions for your AI assistant
Text to copy: paste into the AI chat, and the assistant will guide you step-by-step through the setup.
-
Your AI Assistants
Your own connections with name, redirect address, label (verified, known, or unknown client), "Connected since," and last access. Clicking an entry opens the details. If admin approval is enabled, your own requests also appear here: Waiting for approval, Not approved (with reason, if provided), or Approved – reconnect now.
-
Disconnect
Immediately ends the assistant's access to your content. This only affects your own connection: colleagues using the same assistant remain connected. To reconnect, log in again with the assistant.
Setup per Program#
Replace <your-platform> with your platform's address. The exact connection URL is in the Overview tab and in your profile under Integrations → AI Assistants (MCP).
-
Claude (claude.ai or Claude Desktop)
Settings → Connectors → Add custom connector. URL: https://<your-platform>.kibi.de/api/mcp. Then select "Connect" and log in via the browser.
-
Claude Code
In the terminal, run
claude mcp add --transport http kibi https://<your-platform>.kibi.de/api/mcp, then in Claude Code, call/mcpand start the login. -
Langdock
In the Langdock workspace, under Integrations, add your own MCP server, enter URL https://<your-platform>.kibi.de/api/mcp, select authentication "OAuth 2.0 (Dynamic Client Registration)," then connect and log in via the browser.
-
Microsoft Copilot Studio
In the agent, under Tools, add a new tool of type "Model Context Protocol," enter the server URL, select authentication "OAuth 2.0" with dynamic discovery, then create the connection and log in via the browser.
-
ChatGPT and other programs
Create an MCP server of type "Streamable HTTP" with the connection URL and "OAuth" login.
Logging and Revocation#
-
Logging
Every call from an external assistant is logged.
-
Connected Clients
The Connected Clients tab shows for each client: name and address, whether it is verified or uses an unknown domain, active accesses, and last access. The details show all callback addresses.
-
Revoke
Immediately ends a client's access; all its accesses become invalid. The assistant must reconnect afterward. Without revocation, access is valid for 365 days.
-
Disconnect in profile
Each person can disconnect their own connection in their profile under Integrations → AI Assistants (MCP). Other people using the same assistant remain connected.
Extending Kai with External Tools#
You connect external MCP servers as connections: Administration → Integrations → Third-Party Systems, Connectors tab.
Click Create Connection and select the type MCP Server (Tools for Kai).
Enter name, server URL (HTTPS only), and authentication: None, Bearer Token, or OAuth. For OAuth, first save the Client ID, Client Secret, and permissions, then connect once to the provider via "Connect."
Use Retrieve Tools to see what the server offers.
Specify who can use the tools (roles, groups, or individuals—without selection, everyone who can use Kai) and for which tools Kai should ask in the chat first. Check everything that modifies or sends something.
With Allow Personal Connectors, employees can additionally connect their own MCP servers to Kai for personal use. Such connections are only visible to the respective person.
Warning
Data privacy: Requests to external MCP servers leave your platform. Only connect services that you and your employees trust. Every tool usage is logged, and it is visible in the chat when Kai uses an external tool.
Frequently Asked Questions#
-
Why can't I connect an AI assistant?
Probably none of your roles have the "Connect AI assistants via MCP" permission. In that case, the consent page shows: "Connecting AI assistants is not enabled for your account. Contact your administration," and the AI Assistants (MCP) section is missing from your profile. By default, only System Admins have this permission; the administration can grant it to other roles.
-
What data does an external AI assistant see?
Only what the logged-in person is allowed to see, and only from modules that are active on your platform.
-
Can I block individual tools for external assistants?
No, this option no longer exists. You control access via the users' permissions and active modules—or remove the "Connect AI assistants via MCP" permission from all roles.
-
What does the yellow warning on the consent page mean?
The application is not a known client. Only allow access if you set up the connection yourself. To exclude this entirely, enable "Allow only known clients" under Security.
-
Can I review new AI assistants beforehand?
Yes. Under Integrations → MCP Servers → Security, set the option "Admin approval required" for "New AI Assistants." A new assistant will then only get access after an admin approves it.
-
How do I revoke an assistant's access?
Under Integrations → MCP Servers in the Connected Clients tab, click "Revoke" for the client. Each person can also disconnect their own assistant: in their profile under Integrations → AI Assistants (MCP), click "Disconnect."
-
What does the MCP interface cost?
It is part of Kibi Connect. External services you connect may have their own costs.
-
How do I completely disable MCP?
Remove the "Connect AI assistants via MCP" permission from all roles, including the System Admin role—exceptionally, this is allowed for this permission. Afterward, no AI assistant can find the interface; connected assistants immediately lose access. Warning: "Reset to default" for the System Admin role restores the permission.